Containerized framework for building control performance comparisons: Model predictive control vs deep reinforcement learning control

Abstract

While both model predictive control (MPC) and deep reinforcement learning control (DRL) have shown significant energy cost savings for building systems, there is a lack of in-depth quantitative study on the comparison between the two. One major obstacle is the lack of a holistic evaluation environment for the building community and the control community to integrate their expertise in studying both model-based and learning-based control methods. To address this challenge, this paper presents a scalable containerized software framework for building control performance comparisons, with a special focus on enabling both optimal model-based control and deep learning-based control. The framework provides a standardized building environment for the control community to benchmark different advanced control strategies, and a flexible software architecture for the building community to standardize their own customized building environments. A preliminary performance comparison of MPC and DRL on a single zone building is performed in the case study. Both MPC and DRL can outperform the rule-based baseline controllers in terms of reducing energy cost and maintaining thermal discomfort. DRL can outperform MPC after a long training time with a predefined reward policy.